Title: Adaptive Constraint Guidance Enables Finetuning-Free Diffusion Models for Inorganic Crystal Structure Generation

Abstract: Identifying inorganic crystal structures that possess specific target properties remains a formidable obstacle within the field of materials science. While generative models, particularly advanced diffusion models, hold the potential to capture intricate data distributions and suggest novel, plausible samples, existing AI systems often fail to yield the diversity, originality, and reliability required for high-stakes applications involving experimentally viable materials. To address these limitations, we present a generative machine learning framework utilizing diffusion models enhanced by adaptive constraint guidance. This methodology allows for the integration of user-specified physical and chemical constraints directly into the generation phase. Designed with human experts in mind, the approach ensures practicality and interpretability, facilitating transparent decision-making and expert-led exploration. We further establish a rigorous, multi-step validation pipeline to guarantee the robustness and validity of the proposed candidates. This pipeline leverages graph neural network estimators, trained to attain density functional theory (DFT) level accuracy, alongside convex hull analysis to evaluate thermodynamic stability. We validated our method through case studies involving several classical inorganic compound families. The resulting preliminary findings highlight the framework’s capacity to produce thermodynamically sound crystal structures that adhere to targeted geometric constraints across a variety of inorganic chemical systems.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC